Fault probability monitoring, regulating and controlling system and method

Abstract

The invention discloses a fault probability monitoring, regulation and control system and method, and relates to the technical field of fault monitoring, an acquisition module acquires related parameters to be determined in each period, determines the optimal order of an autoregression model of each related parameter by using an Akaike information criterion, verifies the related parameters to be determined in each period, and determines the optimal order of the autoregression model; whether secondary collection and replacement are carried out or not is decided based on verification results, related parameter vectors of each period are generated, a monitoring module translates windows in a reverse order along the period from the related parameter vectors of each period, comprehensive feature vectors in each window are calculated and sorted, and a core feature vector set is generated through extraction of a time sequence self-encoder; and training the hidden Markov model by adopting a forward-backward algorithm, obtaining the fault probability of each period based on Bayesian reasoning, comparing the fault probability with a probability threshold to judge whether a fault exists or not, and executing regulation and control when the fault exists, thereby realizing high-precision fault monitoring and regulation and control based on acquisition self-inspection, multi-dimensional data fusion and probabilistic reasoning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault monitoring, in particular to a fault probability monitoring and regulation system and method. BACKGROUND

[0002] In the scenarios of industrial production, equipment operation and maintenance, etc., the stable operation of various systems and equipment is of great importance. Once a fault occurs, it may not only lead to production interruption and efficiency reduction, but also cause economic losses, and even endanger personnel safety. Accurate and timely fault monitoring can find potential abnormalities in advance, so as to gain time for preventive maintenance and fault repair, and ensure the reliable operation of the system, so it is necessary to carry out fault monitoring.

[0003] However, the existing fault monitoring technology has the following deficiencies:

[0004] The collection link lacks verification and analysis of collected data, and it is difficult to avoid misjudgment of equipment failure caused by sensor failure;

[0005] The monitoring link does not sufficiently mine multi-dimensional data features, lacks efficient fusion feature extraction and probability reasoning methods, and the accuracy and efficiency of fault identification need to be improved. SUMMARY

[0006] The purpose of the present application is to provide a fault probability monitoring and regulation system and method, to provide a fault monitoring method that realizes sensor collection self-checking, multi-dimensional related data deep mining and probability reasoning, and an adaptive regulation method.

[0007] The technical solution for achieving the purpose of the present application is:

[0008] The fault probability monitoring and regulation system comprises a collection module, a monitoring module and a regulation module;

[0009] The collection module collects M to-be-determined related parameters in the kth period, determines the optimal order q m ) of the autoregressive model Ar(q calculates the mth theoretical related parameter in the kth period verifies the mth to-be-determined related parameter in the kth period based on the verification result, decides whether to perform secondary collection and replacement, and generates the related parameter vector x Fr (k) of the kth period, k=1, 2, 3, …, m=1, …, M;

[0010] The monitoring module obtains the related parameter vector x Fr(k) Start to shift the window in reverse order of the cycle N times, calculate the integrated feature vector in each window, and extract the core feature vector set Y through the time series autoencoder, train the hidden Markov model using the forward-backward algorithm, and get the fault probability p(S k = 2|Y, λ * ) of the kth cycle based on Bayesian inference, compare it with the probability threshold to determine whether there is a fault, where S k = 2 indicates that the state of the kth cycle is a fault, and λ and λ * are the parameters and optimal parameters of the hidden Markov model, respectively.

[0011] The control module stops the operation of the device when the kth cycle is determined to be a fault, and based on the error between each related parameter of the kth cycle and the corresponding standard related parameter, the control amount of the control device corresponding to each related parameter of the kth cycle is converted based on the PID control algorithm and executed.

[0012] Further, the acquisition module includes an acquisition control unit and an acquisition verification unit.

[0013] The acquisition control unit acquires M pending related parameters in the kth cycle to generate a pending related parameter vector Based on the resampling label vector ξ(k) of the kth cycle, determine the related parameters that need to be resampled, and replace the corresponding pending related parameters with the resampling results to generate a related parameter vector x Fr (k).

[0014] The acquisition verification unit extracts the mth pending related parameter of the kth cycle. m Iterate through the value set of order q m , fit the autoregressive model Ar(q m ) corresponding to each value of order q m based on the mth related parameter before the kth cycle, and select the best order q with the smallest Akaike information criterion and the corresponding best autoregressive model Substitute the mth related parameter of the kth cycle with the mth theoretical related parameter of the kth cycle and combine the mth pending related parameter of the kth cycle to calculate the mth parameter error of the kth cycle, compare the mth parameter error of the kth cycle with the mth parameter error threshold, if greater than or equal to the mth parameter error threshold, set the mth resampling label ξ m (k) of the kth cycle to 1, if less than the mth parameter error threshold, set the mth resampling label ξ m(k) set to 0, generate the resampling label vector ξ(k) of the kth period and feed back to the acquisition control unit.

[0015] Further, the acquisition verification unit determines the optimal order of the mth correlation parameter and the corresponding optimal autoregressive model comprising the following steps:

[0016] traverse the value set of the order q m , construct a variational autoregressive model Ar(q m ) based on a single value of the order q m , that is, the mth correlation parameter of each period is equal to the weighted sum of the mth correlation parameters of the q m periods before each period;

[0017] obtain the mth correlation parameters of the q m +1 periods before the kth period, and use the least squares method to fit to obtain the q m autoregressive coefficients of the variational autoregressive model Ar(q m );

[0018] calculate the Akaike information criterion AIC(q m ) of the variational autoregressive model Ar(q m ) of the single value of the order q m ; select the optimal order q with the minimum Akaike information criterion and obtain the corresponding optimal autoregressive model

[0019] Further, the monitoring module comprises a feature processing unit and a fault determination unit.

[0020] The feature processing unit initializes the end period of the window as the kth period and shifts it by a single period in reverse order q times, calculates the time series statistics and cross-correlation coefficient matrix in each window and flattens it into a vector, splices it with the correlation parameter vector of the end period in each window to generate the corresponding comprehensive feature vector, retains the fault sensitive information of the comprehensive feature vector of each window through the time series autoencoder and captures the period correlation, and generates a core feature vector set Y.

[0021] The fault determination unit defines a state S, uses the forward-backward algorithm to maximize the log-likelihood function logp(Y|λ) as the target, combines the core feature vector set Y to train the hidden Markov model to obtain the optimal parameter set λ * , and calculates the fault probability p(S k =2|Y,λ *) and compared with a probability threshold, if greater than or equal to the probability threshold, it is determined that there is a failure in the kth period, if less than the probability threshold, no processing is performed, wherein λ is a parameter set of the hidden Markov model.

[0022] Specifically, the nth window includes the relevant parameter vectors of the k-n-N0+2th period to the k-n+1th period, N0 is the window size, the mth relevant parameter of the N0 periods in the nth window is obtained, the window mean μ n,m , the window standard deviation σ n,m , the window maximum value , the window minimum value , the window skewness Sk n,m , the window kurtosis Ku n,m and the lag one autocorrelation Af n,m , wherein the window skewness Sk n,m , the window kurtosis Ku n,m and the lag one autocorrelation Af n,m are specifically as follows:

[0023]

[0024]

[0025] , the window skewness Sk n,m and the window kurtosis Ku n,m respectively reflect the symmetry and tail thickness of the numerical distribution of the mth relevant parameter in the nth window, and the lag one autocorrelation Af n,m is used to measure the time series linear correlation degree of the mth relevant parameter in the nth window, the relevant parameters in the nth window are combined in pairs, the covariance of each combination is calculated and divided by the product of the window standard deviations of the two relevant parameters in the combination, the cross-correlation coefficient of each combination is obtained and arranged into the cross-correlation coefficient matrix ρ n of the nth window, the upper triangular elements in the cross-correlation coefficient matrix ρ n of the nth window are extracted and spliced with the time series statistics of the M relevant parameters of the nth window and the relevant parameter vector of the k-n+1th period to generate the comprehensive feature vector of the nth window through Z-Score standardization processing to obtain a comprehensive feature vector set

[0026] Further, the time series autoencoder includes a bidirectional encoding layer, an attention aggregation layer and a reconstruction verification layer.

[0027] The bidirectional encoding layer uses forward and backward long short-term memory networks to process the comprehensive feature vector set The forward long short-term memory network transforms the comprehensive feature vector of the nth window into The forward cell state of the n+1th window With the forward hidden state Mapped to the forward cell state of the nth window With the forward hidden state Loop until the forward cell state of the first window is generated and the forward hidden state The backward long short-term memory network starts from the first window in reverse order until the backward cell state of the Nth window is generated. and the backward hidden state Stop when n is reached and concatenate the forward hidden state of the nth window and the backward hidden state Generate the corresponding hidden state h n , organized into a hidden state sequence

[0028] The attention cohesion layer takes the hidden state h of the nth window n The linear modulation of is mapped by the Tanh function and multiplied by the attention vector to obtain the attention score of the nth window And normalized to the attention weight of the nth window through the Softmax function Multiply the attention weight of each window with the hidden state and sum them to generate the context hidden state h Ag And concatenate it with the hidden state of each window respectively, map it to the core feature vector of each window through linear modulation and ReLU function, and generate the core feature vector set Y;

[0029] The reconstruction validation layer converts the core feature vector y of the nth window in the core feature vector set Y into n Generate the reconstructed comprehensive feature vector of the nth window through linear modulation and ReLU function mapping And organize it into a reconstructed comprehensive feature vector set Compute and reconstruct comprehensive feature vector set and comprehensive feature vector set The reconstruction error is compared with the reconstruction error threshold. If and only if the reconstruction error is greater than or equal to the reconstruction error threshold, the parameters of the feature embedding layer, the bidirectional encoding layer, and the attention cohesion layer are reversely updated based on the reconstruction error threshold and the comprehensive feature vector set is reprocessed.

[0030] Furthermore, the failure probability p(S k =2|Y,λ * ) comprises the following steps:

[0031] Define S=1 and S=2 to indicate that the state S is normal and faulty respectively, initialize the parameter set λ, including the state probability vector π of the Nth window N , transfer matrix A and observation probability matrix B, where the j1th row and j2th column A(j1, j2) of the state transfer matrix A reflects the probability of state S = j1 transferring to state S = j2, and the j1th row and nth column B(j1, n) of the observation probability matrix B represents the core feature vector of the nth window appearing in state S = j1. The probability of j1, j2∈{1,2};

[0032] Define the state S of the nth window k-n+1 =j1 and the probability of observing the core feature vector from the Nth window to the n+1th window is the forward variable α n (j1), forward variable α n (j1) satisfies the forward recursive formula, as follows:

[0033]

[0034] Establish the forward variable α according to the forward recursive formula n (j1) and the forward variable α N (j3) linear association, forward variable α N (j3) The details are as follows:

[0035] α N (j3) = π N (S k-N+1 =j3)·B(j3,N),

[0036] Among them, π N (S k-N+1 =j3) is the state probability vector π of the Nth window N Medium state S k-N+1 = the probability of j3, j3∈{1,2};

[0037] Define the state S of the nth window k-n+1 =j1 and the probability of observing the core feature vector from the n-1th window to the 1st window is the backward variable β n (j1), backward variable β n (j1) satisfies the backward recursive formula, as follows:

[0038]

[0039] Establish the backward variable β according to the backward recursion formula n (j1) is linearly associated with the backward variable α1(j3) = 1;

[0040] Define the state S of the nth windowk-n+1 = probability of single state probability γ n (j1), specifically as follows:

[0041]

[0042] single state probability γ n (j1) reflects the expectation of the state S k-n+1 of the n-th window;

[0043] define the state S k-n+1 of the n-th window = j1 and the state S k-n+2 of the (n-1)-th window = j2 as a probability of double state joint probability γ n (j1, j2), specifically as follows:

[0044]

[0045] double state joint probability γ n (j1, j2) reflects the expectation of the state S k-n+1 of the (n-1)-th window from the state S k-n+2 of the n-th window;

[0046] In order to maximize the log-likelihood function logp(Y|λ), the parameter set λ is updated in each iteration until the parameter set changes of adjacent iterations are less than a convergence threshold, and the optimal parameter set λ is obtained * wherein p(Y|λ) = α1(1) + α1(2) = β N (1) + β N (2);

[0047] In combination with the optimal state probability vector π * of the N-th window in the optimal parameter set λ * N , the optimal state transition matrix A * and the optimal observation probability matrix B * are used for forward recursion to obtain the fault probability p(S k = 2|Y, λ * ) of the k-th period.

[0048] Further, the updating of the parameter set λ of the hidden Markov model in the (l+1)-th round includes:

[0049] The probability of the state S k-N+1 = j1 in the state probability vector of the N-th window in the (l+1)-th round is updated to the single state probability calculated according to the parameter set λ l of the l-th round;

[0050] The state transfer matrix A of the l+1th round l+1 A in row j1, column j2 l+1 (j1, j2) is updated to the parameter set λ according to the first round l The ratio of the expected number of times that state S=j1 transitions to state S=j2 to the total expected number of times state S=j1 is calculated is as follows:

[0051]

[0052] in, and are respectively based on the parameter set λ of the first round l Calculate the two-state joint probability and single-state probability of the nth window;

[0053] Update the observation probability matrix B of the l+1th round based on the maximum likelihood estimation of Gaussian distribution l+1 B in row j1, column n l +1 (j1,n), as follows:

[0054] B l+1 (j1,n)=Gauss(μ l (j1,n),Σ l (j1,n)),

[0055] Among them, Gauss represents Gaussian distribution, μ l (j1,n) and Σ l (j1,n) are the parameter set λ according to the first round l Calculate the state S of the nth window k-n+1 = the mean and variance of the distribution of j1. The specific calculation formula is as follows:

[0056]

[0057] The fault probability monitoring and control method includes the following steps:

[0058] In the kth cycle, the relevant parameter vector to be determined is acquired Determine the optimal order that minimizes the Akaike Information Criterion based on the mth correlation parameter fitting before the kth period and the corresponding optimal autoregressive model

[0059] Before the kth cycle Substitute the mth related parameter of the period into the best autoregressive model Get the mth theoretical related parameters of the kth cycle Combined with the mth undetermined related parameter of the kth period calculating the mth parameter error of the kth cycle and comparing it with the mth parameter error threshold value;

[0060] if greater than or equal to the mth parameter error threshold value, performing secondary acquisition and taking the acquisition result as the mth related parameter of the kth cycle if less than the mth parameter error threshold value, directly taking the mth pending related parameter of the kth cycle as the mth related parameter of the kth cycle generating a related parameter vector x Fr (k);

[0061] translating the window from the kth cycle along the cycle in reverse order by one cycle N times, calculating the time series statistics and cross-correlation coefficient matrix in each window and flattening them into a vector, splicing the related parameter vector of the end cycle in each window into a comprehensive feature vector in each window, and using a time series autoencoder to extract a core feature vector set Y;

[0062] defining a state S and using a forward-backward algorithm to maximize a log-likelihood function logp(Y|λ) as an objective, combining the core feature vector set Y to train the parameters λ of the hidden Markov model to obtain an optimal parameter set λ * using Bayesian inference to calculate the fault probability p(S k =2|Y, λ * ) of the kth cycle and comparing it with a probability threshold value to determine whether the kth cycle is faulty, wherein S k =2 indicates that the state of the kth cycle is faulty;

[0063] if greater than or equal to the probability threshold value, determining that a fault occurs in the kth cycle and stopping the operation of the device, converting the error of each related parameter of the kth cycle and the corresponding standard related parameter into a control amount of the control device corresponding to each related parameter of the kth cycle based on a PID control algorithm and performing control, and if less than the probability threshold value, not doing anything.

[0064] Compared with the prior art, the acquisition module acquires the pending related parameters in each cycle, determines the optimal order of the autoregressive model of each related parameter using the Akaike information criterion and verifies the pending related parameters of each cycle, decides whether to perform secondary acquisition and replacement based on the verification result, generates a related parameter vector of each cycle, the monitoring module translates the window from the related parameter vector of each cycle along the cycle in reverse order, calculates and arranges the comprehensive feature vector in each window, and extracts a core feature vector set through a time series autoencoder, trains a hidden Markov model using a forward-backward algorithm and obtains the fault probability of each cycle based on Bayesian inference, compares it with a probability threshold value to determine whether a fault exists, and realizes high-precision fault monitoring based on acquisition self-checking, multi-dimensional data fusion and probabilistic reasoning. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 This is a schematic diagram of the fault probability monitoring and control system;

[0066] Figure 2 Flowchart for determining the optimal order of autoregressive model;

[0067] Figure 3 This is the processing flow chart of the time series self-encoder;

[0068] Figure 4 This is a flow chart of the fault probability monitoring and control method. DETAILED DESCRIPTION

[0069] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0070] Example 1:

[0071] like Figure 1 As shown, a specific embodiment of the present invention discloses a fault probability monitoring and control system, including an acquisition module, a monitoring module and a control module;

[0072] The acquisition module collects the relevant parameters to be determined through M main sensors in the kth cycle, and uses the Akaike information criterion to determine the autoregressive model Ar(q m ) Before combination The mth related parameter of the kth period is used to deduce the mth theoretical related parameter of the kth period. And verify the mth undetermined related parameter in the kth cycle Based on the verification results, decide whether to enable the corresponding secondary sensor for secondary collection and replacement, and generate the relevant parameter vector x of the kth cycle Fr (k), k=1,2,3,…,m=1,…,M;

[0073] The monitoring module obtains the relevant parameter vector x from the kth cycle Fr (k), the window is shifted N times in reverse order along the cycle, the comprehensive feature vector in each window is calculated and sorted, and the fault-sensitive information and cycle association are retained by the time series autoencoder to generate the core feature vector set Y. The forward-backward algorithm is used to maximize the log-likelihood function logp(Y|λ) to train the hidden Markov model. Combined with Bayesian reasoning, the failure probability p(S) of the kth cycle is obtained. k =2|Y,λ * ) and compared with the probability threshold to determine whether there is a fault, where S k =2 indicates that the state of the kth cycle is faulty, λ and λ *are the parameters and optimal parameters of the hidden Markov model, respectively. p(Y|λ) represents the probability of the core feature vector set Y appearing under the condition of parameter λ.

[0074] The control module stops the equipment operation when it is judged as a fault in the kth cycle. According to the error between each relevant parameter of the kth cycle and the corresponding standard relevant parameter, it is converted into the control amount of the control device corresponding to each relevant parameter of the kth cycle based on the existing PID control algorithm and the control is performed.

[0075] Furthermore, the acquisition module includes an acquisition control unit and an acquisition verification unit;

[0076] The acquisition control unit collects different undetermined related parameters through M main sensors in the kth cycle and obtains the undetermined related parameter vector Based on the re-sampling annotation vector of the kth period ξ(k)=[ξ1(k),…,ξ m (k),…,ξ M (k)] Select the secondary sensor to perform secondary collection in the kth cycle, and prompt the staff to replace or repair the corresponding main sensor when the equipment is idle, and replace the undetermined related parameter vector with the secondary collection result The corresponding undetermined related parameters in the , the unreplaced undetermined related parameters are regarded as undetermined related parameters, and the related parameter vector is generated in, and are the mth undetermined related parameters and related parameters of the kth period, ξ m (k) is the mth re-collection mark of the kth period, when ξ m When (k) = 1, it means that the corresponding secondary sensor is enabled in the kth cycle to replace the faulty primary sensor. m When (k) = 0, it means that the secondary sensor does not need to be enabled in the kth cycle, m = 1,…,M, and the relevant parameters are the operating parameters closely related to the fault in the actual operation of the equipment;

[0077] The acquisition verification unit collects relevant parameter vectors from the undetermined Extract the mth undetermined related parameters of the kth period in sequence Ergodic order q m The value set of the m-th related parameter vector before the k-th period is used to fit the order q m Each value of the autoregressive model Ar(q m ) and select the optimal order with the smallest Akaike Information Criterion and the corresponding optimal autoregressive model Before the kth cycle Substitute the mth related parameter of the period into the best autoregressive model get the mth theoretical correlation parameter of the kth cycle combine the mth undetermined correlation parameter of the kth cycle calculate the mth parameter error of the kth cycle, determine whether the mth parameter error of the kth cycle is greater than or equal to the mth parameter error threshold, if greater than or equal to the mth parameter error threshold, set the mth reacquisition flag of the kth cycle as 1 m (k) to 1, if less than the mth parameter error threshold, set the mth reacquisition flag of the kth cycle as 0 m (k) to 0, arrange the reacquisition flags of the M correlation parameters of the kth cycle as the reacquisition flag vector of the kth cycle ξ(k) and feed back to the acquisition control unit.

[0078] As shown in Figure 2 , further, the acquisition verification unit determines the optimal order of the mth correlation parameter and the corresponding optimal autoregressive model comprising the following steps:

[0079] regard the order q m of the mth correlation parameter as a variable and traverse the value set of the order q m ;

[0080] construct a variational autoregressive model Ar(q m ) based on a single value of the order q m , that is, the mth correlation parameter of each cycle is equal to the weighted sum of the mth correlation parameters of the q m preceding cycles of each cycle;

[0081] call the mth correlation parameters of the q m +1 preceding cycles of the kth cycle, adjust the q m autoregressive coefficients of the variational autoregressive model Ar(q m ) to minimize the error between the weighted sum of the mth correlation parameters of the q m preceding cycles and the mth correlation parameter x m (k-1) of the k-1th cycle by using the least square method, and fit to obtain the autoregressive coefficient vector of the variational autoregressive model Ar(q m );

[0082] calculate the Akaike information criterion AIC(q m ) of the variational autoregressive model Ar(q m ) of a single value of the order q m , the specific formula is as follows:

[0083]

[0084] wherein, the k-1th period based on a variational autoregressive model Ar(q m ) calculated;

[0085] selecting the optimal order with the minimum Akaike information criterion obtaining a corresponding optimal autoregressive model

[0086] Further, the monitoring module comprises a feature processing unit and a fault determination unit.

[0087] The feature processing unit initializes the end period of the window as the kth period and shifts it in a reverse order of periods N times, each time moving forward by one period, calculates the time series statistics and the cross-correlation coefficient matrix in each window and flattens them into a vector, splices the correlation parameter vector of the end period in each window to generate a comprehensive feature vector set in each window, retains the fault sensitive information of the comprehensive feature vector set in each window through the time series autoencoder and captures the period correlation to generate a core feature vector set

[0088] The fault determination unit defines a state S, uses the forward-backward algorithm to maximize the log-likelihood function logp(Y|λ) as the target, and trains the hidden Markov model to obtain the optimal parameter set λ combining the core feature vector set Y * , calculates the fault probability p(S k =2|Y,λ * ) of the kth period using Bayesian inference and compares it with the probability threshold, if greater than or equal to the probability threshold, it is determined that the device has a fault in the kth period, if less than the probability threshold, it does not do anything.

[0089] Specifically, the end period of the nth window is the k-n+1th period, since the size of the window is fixed as N0, the nth window includes the correlation parameter vectors from the k-n-N0+2th period to the k-n+1th period, the mth correlation parameter of the N0 periods in the nth window is obtained, the time series statistics of the mth correlation parameter in the nth window are calculated respectively, including the window mean μ n,m , the window standard deviation σ n,m , the window maximum value the window minimum value the window skewness Sk n,m , the window kurtosis Ku n,m and the lag-1 autocorrelation Af n,m , wherein the specific formulas of the window skewness Sk n,m , the window kurtosis Ku n,m and the lag-1 autocorrelation Af n,m are as follows:

[0090]

[0091] Among them, the window skewness Sk n,m and window kurtosis Ku n,m Respectively reflect the symmetry and tail thickness of the numerical distribution of the mth correlation parameter in the nth window, the lag-one autocorrelation Af n,m Used to measure the linear correlation of the mth related parameter in the nth window in time series, and calculate the m1th related parameter in the nth window and the m2th related parameter The covariance of And divided by the m1th related parameter The window standard deviation Related parameters for m2 The window standard deviation The product of the two, we get the cross-correlation coefficient matrix ρ of the nth window n The correlation coefficient of the m1th row and m2th column in From the cross-correlation matrix ρ of the nth window n Extract the upper triangular elements containing the diagonal elements, the time series statistics of the M related parameters in the nth window, and the related parameter vector of the k-n+1th period Splice by column and generate the comprehensive feature vector of the nth window through Z-Score normalization Arrange the comprehensive feature vectors of N windows in periodic order to obtain the comprehensive feature vector set Since the windows are shifted in reverse order, the end period corresponding to the Nth window is the earliest.

[0092] like Figure 3 As shown, further, the temporal autoencoder includes a bidirectional encoding layer, an attention aggregation layer, and a reconstruction verification layer;

[0093] The bidirectional encoding layer uses parallel forward long short-term memory networks and backward long short-term memory networks to process the comprehensive feature vector set respectively. The forward long short-term memory network uses the forget gate to discard the forward cell state of the n+1th window according to the cycle order. Part of the information in the n-th window is obtained by using the input gate to obtain the comprehensive feature vector and the forward hidden state of the n+1th window Select some information and map it to the forward cell state of the nth window Use the output gate to concatenate the processing results of the forget gate and the input gate and map them to generate the forward hidden state of the nth window The cycle repeats until the forward cell state of the first window is generated and the forward hidden state The loop stops when the Nth window is reached, and the backward long short-term memory network starts from the first window in reverse order of the cycle, and uses the forget gate, input gate, and output gate to repeat the process until the backward cell state of the Nth window is generated. and the backward hidden state Stop when , concatenate the forward hidden state and backward hidden state of the same window to generate the hidden state of N windows, and arrange them in periodic order as a hidden state sequence

[0094] The attention cohesion layer transforms the hidden state sequence The hidden state h of the nth window in n After linear modulation and Tanh function mapping to the range of [-1,1] and multiplying with the preset attention vector, the attention score of the nth window is obtained. Arranged in order of cycles as attention score vector s At And normalized to the attention weight vector w through the Softmax function At , the attention weight vector w At With hidden state sequence The attention weight of each window in is multiplied by the hidden state and summed to generate the context hidden state h Ag , the context hidden state h Ag Spliced ​​with the hidden state of each window and mapped to the core feature vector of each window through linear modulation and ReLU function to generate a core feature vector set

[0095] The reconstruction validation layer converts the core feature vector y of the nth window in the core feature vector set Y into n By expanding the dimension through linear modulation and enhancing the nonlinear expression through the ReLU function, the reconstructed comprehensive feature vector of the nth window is generated by mapping Generate and reconstruct comprehensive feature vector set Compute and reconstruct comprehensive feature vector set and comprehensive feature vector set The reconstruction error is compared with the reconstruction error threshold. If the reconstruction error is less than the reconstruction error threshold, the core feature vector set Y is determined to be qualified. If the reconstruction error is greater than or equal to the reconstruction error threshold, the core feature vector set Y is determined to be over-condensed and the comprehensive feature vector set is lost. The fault-sensitive information in the core feature vector set Y cannot be effectively restored. Based on the reconstruction error threshold, the parameters of the feature embedding layer, bidirectional encoding layer and attention cohesion layer are reversely updated and the comprehensive feature vector set is reprocessed.

[0096] Furthermore, the failure probability p(S k =2|Y,λ *) comprises the following steps:

[0097] Define state S = {1, 2}, where S = 1 and S = 2 represent the state S is normal and faulty respectively;

[0098] Initialize the state probability vector π of the Nth window N , state probability vector π N Including the probabilities when the state S is normal and faulty, where the Nth window corresponds to the k-N+1th cycle, which is the earliest cycle in the cycle sequence;

[0099] Initialize the state transfer matrix A and the observation probability matrix B, where the j1th row and j2th column A(j1, j2) of the state transfer matrix A reflects the probability of transitioning from state S = j1 to state S = j2, and the j1th row and nth column B(j1, n) of the observation probability matrix B represents the core feature vector of the nth window appearing in state S = j1. The probability of j1, j2∈{1,2};

[0100] The parameter set λ that defines the hidden Markov model includes the state probability vector π of the Nth window N , state transfer matrix A and observation probability matrix B;

[0101] Define the state S of the nth window k-n+1 =j1 and the probability of observing the core feature vector from the Nth window to the n+1th window is the forward variable α n (j1), and the forward variable α n (j1) satisfies the forward recursive formula, as follows:

[0102]

[0103] According to the forward recursive formula, the forward variable α n (j1) can be organized into forward variables α N (j3) is a linear combination of the elements in the state transfer matrix A and the observation probability matrix B, and the forward variable α N (j3) The details are as follows:

[0104] α N (j3) = π N (S k-N+1 =j3)·B(j3,N),

[0105] Among them, S k-N+1 =j3 reflects the state of the Nth window at the beginning of the cycle sequence, π N (S k-N+1 =j3) is the state probability vector π of the Nth window N Medium state S k-N+1= the probability of j3, j3∈{1,2};

[0106] Define the state S of the nth window k-n+1 =j1 and the probability of observing the core feature vector from the n-1th window to the 1st window is the backward variable β n (j1), and the backward variable β n (j1) satisfies the backward recursive formula, as follows:

[0107]

[0108] According to the backward recursive formula, the backward variable β n (j1) can be organized into a linear combination of the backward variable α1(j3) and the elements in the state transfer matrix A and the observation probability matrix B. The backward variable α1(j3) = 1, j3∈{1,2}. This is because the kth cycle is the end cycle in the cycle sequence, and the state and core feature vector of each window have been determined;

[0109] Define the state S of the nth window k-n+1 =j1 probability is the single state probability γ n (j1), as follows:

[0110]

[0111] Single state probability γ n (j1) reflects the state S of the nth window k-n+1 expectations;

[0112] Define the state S of the nth window k-n+1 = j1 and the state S of the n-1th window k-n+2 =j2 is the joint probability of the two states γ n (j1,j2), as follows:

[0113]

[0114] Two-state joint probability γ n (j1, j2) reflects the state S from the nth window k-n+1 Transfer to the state S of the n-1th window k-n+2 expectations;

[0115] With the goal of maximizing the log-likelihood function logp(Y|λ), the parameter set λ of the hidden Markov model is updated in each round of iteration until the change in the parameter set of the adjacent iteration is less than the convergence threshold, and the parameter set λ of the last round is taken as the optimal parameter set λ. * , where, p(Y|λ)=α1(1)+α1(2)=β N (1)+βN (2);

[0116] Combined with the optimal parameter set λ * The optimal state probability vector π of the Nth window in * N , using the optimal state transfer matrix A * and the optimal observation probability matrix B * Perform forward recursion to obtain the failure probability p(S k =2|Y,λ * ).

[0117] Furthermore, the update of the parameter set λ of the hidden Markov model in the l+1th round includes:

[0118] The state probability vector of the Nth window in the l+1th round Medium state S k-N+1 = the probability of j1 Update to the parameter set λ based on the first round l Calculated single-state probability

[0119] The state transfer matrix A of the l+1th round l+1 A in row j1, column j2 l+1 (j1, j2) is updated to the parameter set λ according to the first round l The ratio of the expected number of transitions from state S=j1 to state S=j2 to the total expected number of transitions from state S=j1 is calculated as follows:

[0120]

[0121] in, and are respectively based on the parameter set λ of the first round l Calculate the two-state joint probability and single-state probability of the nth window;

[0122] Update the observation probability matrix B of the l+1th round based on the maximum likelihood estimation of Gaussian distribution l+1 B in row j1, column n l +1 (j1,n), as follows:

[0123] B l+1 (j1,n)=Gauss(μ l (j1,n),Σ l (j1,n)),

[0124] Among them, Gauss represents Gaussian distribution, μ l (j1,n) and Σ l(j1,n) are the parameter set λ according to the first round l Calculate the state S of the nth window k-n+1 = the mean and variance of the distribution of j1. The specific calculation formula is as follows:

[0125]

[0126] Example 2

[0127] like Figure 4 As shown, a specific embodiment of the present invention discloses a fault probability monitoring and control method, which is executed based on the fault probability monitoring and control system and includes the following steps:

[0128] In the kth cycle, different undetermined related parameters are collected by M main sensors respectively to obtain the undetermined related parameter vector Determine the optimal order that minimizes the Akaike Information Criterion based on the mth correlation parameter fitting before the kth period and the corresponding optimal autoregressive model

[0129] The kth period before Substitute the mth related parameter of the period into the best autoregressive model Get the mth theoretical related parameters of the kth cycle Combined with the mth undetermined related parameter of the kth period Calculate the mth parameter error of the kth cycle and compare it with the mth parameter error threshold;

[0130] If it is greater than or equal to the mth parameter error threshold, the secondary sensor is enabled to perform secondary collection in the kth cycle and the collection result is used as the mth related parameter of the kth cycle. If it is less than the mth parameter error threshold, the mth undetermined related parameter of the kth period is directly used as the mth related parameter of the kth period. Generate relevant parameter vector x Fr (k);

[0131] The window is shifted N times in reverse order starting from the kth cycle. The time series statistics and cross-correlation coefficient matrix within each window are calculated and flattened into a vector. The vectors are then concatenated with the relevant parameter vectors of the end cycle within each window to form a comprehensive feature vector within each window. A time series autoencoder is used to retain the fault-sensitive information of the comprehensive feature vector of each window and capture cycle correlations to generate the core feature vector set Y.

[0132] Define the state S and use the forward-backward algorithm to maximize the log-likelihood function logp(Y|λ) as the goal, and combine the core feature vector set Y to train the parameters λ of the hidden Markov model to obtain the optimal parameter set λ* , the failure probability p(S k = 2 | Y, l * ) is calculated using Bayesian inference and compared with a probability threshold, where S k = 2 indicates that the state of the kth cycle is failure.

[0133] If greater than or equal to the probability threshold, it is determined that the device fails in the kth cycle and stops the operation of the device, and according to the error of each related parameter of the kth cycle and the corresponding standard related parameter, the PID control algorithm is converted into the control amount of the control device corresponding to each related parameter of the kth cycle and executed, and if less than the probability threshold, no action is taken.

[0134] Embodiment 3:

[0135] As an example, the failure probability monitoring and control system is connected with the pressure sensor deployed at the filter material of the intelligent filter through the PLC controller, the differential pressure of the filter material to be determined is collected through the main pressure sensor in the kth cycle, the best autoregressive model is determined and verified for fitting respectively to decide whether to enable the secondary pressure sensor to re-collect, the integrated feature vector of the filter material differential pressure of the kth cycle and the previous N-1 cycles is calculated using the window integration of the reverse translation, the core feature vector set Y is generated by processing using the time series autoencoder, the hidden Markov model is trained and the probability of filter material blockage in the kth cycle is calculated based on Bayesian inference, when it is determined that the kth cycle appears blockage, the cleaning flow rate and cleaning time of the cleaning component are determined according to the filter material differential pressure and the standard filter material differential pressure without blockage through the PID control algorithm, that is, when applied to the filter material of the intelligent filter, the related parameter is the filter material differential pressure, the failure is the filter material blockage, the control device is the cleaning device, and the control amount includes the cleaning flow rate and the cleaning time.

[0136] Embodiment 4:

[0137] As an example, the failure probability monitoring and regulation system is connected with the temperature sensor and the rotating speed sensor arranged in the intelligent filter through the PLC controller, the standby temperature and the standby motor rotating speed are collected through the main temperature sensor and the main rotating speed sensor in the kth period, the best autoregressive model is determined and verified respectively to decide whether to start the corresponding secondary sensor to re-collect, the temperature and the motor rotating speed of the kth period and the N-1 periods before are integrated and calculated by using the window of the reverse translation, the comprehensive feature vector is obtained, the core feature vector set Y is generated by using the time sequence auto-encoder processing, the hidden Markov model is trained, and the probability of the abnormal operation of the intelligent filter in the kth period is calculated based on the Bayesian inference, if it is judged to be abnormal, the power adjustment amount of the temperature control equipment and the power adjustment amount of the motor are determined respectively by using the PID control algorithm according to the error of the temperature in the kth period and the standard temperature and the error of the motor rotating speed and the standard motor rotating speed, that is, when applied to the intelligent filter, the related parameters include the temperature and the motor rotating speed, the failure is the operation abnormality, the regulation equipment is the temperature control equipment and the motor, and the regulation amount includes the power adjustment amount of the temperature control equipment and the motor.

[0138] The application discloses a failure probability monitoring and regulation system and method, which comprises a collection module, a monitoring module and a regulation module, the collection module collects standby related parameters in each period, determines the best order of the autoregressive model of each related parameter by using the Akaike information criterion, verifies the standby related parameters in each period, decides whether to perform secondary collection and replacement based on the verification result, and generates a related parameter vector of each period, the monitoring module starts from the related parameter vector of each period, translates the window in reverse order along the period, calculates and arranges the comprehensive feature vector in each window, extracts and generates a core feature vector set through a time sequence auto-encoder, trains a hidden Markov model by using a forward-backward algorithm, and obtains the failure probability of each period based on Bayesian inference, compares the probability threshold to determine whether there is a failure, and the regulation module enables corresponding regulation equipment to perform regulation based on a PID control algorithm when there is a failure, so that high-precision failure monitoring and regulation based on collection self-checking, multi-dimensional data fusion and probability inference are realized.

[0139] The above only describes the preferred embodiments of the application, and the protection scope of the application is not limited to the above-described embodiments, and any technical scheme falling within the idea of the application belongs to the protection scope of the application. It should be noted that, for ordinary skilled persons in the technical field, some improvements and decorations without departing from the principle of the application are also regarded as the protection scope of the application.

Claims

1. A fault probability monitoring and control system, comprising a control module, which stops the operation of the equipment when a fault is determined and performs control based on a PID control algorithm, characterized in that: It also includes an acquisition module and a monitoring module; The acquisition module acquires M undetermined related parameters in the kth period, determines the optimal order of the autoregressive model of the mth related parameter using the Akaike information criterion, infers the mth theoretical related parameter of the kth period and verifies the mth undetermined related parameter of the kth period, decides whether to perform secondary acquisition and replacement based on the verification result, and generates a related parameter vector of the kth period, where k=1, 2, 3, ..., and m=1, ..., M; The monitoring module shifts the window N times in reverse order along the cycle starting from the relevant parameter vector of the kth cycle, calculates and organizes the comprehensive feature vectors in each window, and generates a core feature vector set through a time series autoencoder. The forward-backward algorithm is used to train the hidden Markov model and the fault probability of the kth cycle is obtained based on Bayesian reasoning, and is compared with the probability threshold to determine whether a fault exists.

2. The fault probability monitoring and control system according to claim 1, characterized in that: The monitoring module includes a feature processing unit and a fault determination unit; The feature processing unit initializes the end period of the window to the kth period and shifts a single period N times in reverse order of the period. It calculates the time series statistics and cross-correlation coefficient matrix within each window and flattens them into a vector. It concatenates them with the relevant parameter vector of the end period within each window to generate a corresponding comprehensive feature vector. The time series autoencoder retains the fault-sensitive information of the comprehensive feature vector of each window and captures the period correlation to generate a core feature vector set. The fault judgment unit defines a state, uses a forward-backward algorithm to maximize the log-likelihood function, combines a core feature vector set to train a hidden Markov model to obtain an optimal parameter set, uses Bayesian reasoning to calculate the failure probability of the k-th cycle and compares it with a probability threshold, and determines that a fault exists in the k-th cycle if and only if the failure probability of the k-th cycle is greater than or equal to the probability threshold.

3. The fault probability monitoring and control system according to claim 2, characterized in that: Calculating the probability of failure in the kth cycle involves the following steps: Define state S = {1, 2} and initialize the parameter set, including the state probability vector, state transition matrix, and observation probability matrix of the Nth window. The j1th row and j2th column of the state transition matrix reflect the probability of transitioning from state S = j1 to state S = j2. The j1th row and nth column of the observation probability matrix represent the probability of the core feature vector of the nth window appearing in state S = j1, j1, j2∈{1, 2}. Define the forward variable α n (j1) is the state S of the nth window k-n+1 =j1 and the probability of observing the core feature vectors from the Nth window to the n+1th window, the forward variable α is established according to the forward recursive formula n (j1) and the forward variable α N (j3) linear association, j3∈{1,2}; Define the backward variable β n (j1) is the state S of the nth window k-n+1 =j1 and the probability of observing the core feature vector from the n-1th window to the 1st window, the backward variable β is established according to the backward recursive formula n (j1) is linearly associated with the backward variable α1(j3) = 1; Define the single-state probability γ n (j1) is the state S of the nth window k-n+1 = the probability of j1; Define the two-state joint probability γ n (j1, j2) is the state S of the nth window k-n+1 = j1 and the state S of the n-1th window k-n+2 = the probability of j2.

4. The fault probability monitoring and control system according to claim 2, characterized in that: Calculating the failure probability of the kth cycle also includes the following steps: With the goal of maximizing the log-likelihood function, in the l+1th round of iteration, the Nth window state S k-N+1 =j1 probability is updated to the single state probability calculated based on the parameter set of the first round Update the j1th row and j2th column of the state transition matrix of the l+1th round to the ratio of the expected number of times the state S=j1 transfers to the state S=j2 calculated based on the parameter set of the lth round to the total expected number of times the state S=j1 transfers; Update the j1th row and nth column of the observation probability matrix of the l+1th round based on the maximum likelihood estimation of the Gaussian distribution; Iterative update stops when the parameter set change between adjacent iterations is less than the convergence threshold, and the optimal parameter set is obtained; Combined with the optimal state probability vector of the Nth window in the optimal parameter set, the optimal state transfer matrix and the optimal observation probability matrix are used for forward recursion to obtain the failure probability of the kth cycle.

5. The fault probability monitoring and control system according to claim 2, characterized in that: The temporal autoencoder includes a bidirectional encoding layer and an attention cohesion layer; The bidirectional encoding layer uses forward and backward long short-term memory networks to process the comprehensive feature vector set respectively. The forward long short-term memory network maps the comprehensive feature vector of the nth window, the forward cell state and the forward hidden state of the n+1th window to the forward cell state and forward hidden state of the nth window. The loop process stops when the forward cell state and forward hidden state of the first window are generated. The backward long short-term memory network loops from the first window in reverse order until the backward cell state and backward hidden state of the Nth window are generated. The forward hidden state and backward hidden state of the nth window are spliced ​​to generate the corresponding hidden state, which is organized into a hidden state sequence. The attention cohesion layer maps the linear modulation of the hidden state of the nth window through the Tanh function and multiplies it with the attention vector to obtain the attention score of the nth window, and normalizes it to the attention weight of the nth window through the Softmax function. The attention weight of each window is multiplied by the hidden state and summed to generate the context hidden state and concatenate it with the hidden state of each window respectively. The core feature vector of each window is mapped through linear modulation and ReLU function to generate a core feature vector set.

6. The fault probability monitoring and control system according to claim 5, characterized in that: The temporal autoencoder further includes a reconstruction verification layer; The reconstruction verification layer generates a reconstructed comprehensive feature vector of the nth window in the core feature vector set through linear modulation and ReLU function mapping, and organizes it into a reconstructed comprehensive feature vector set, calculates the reconstruction error between the reconstructed comprehensive feature vector set and the comprehensive feature vector set, and compares it with the reconstruction error threshold. If and only if the reconstruction error is greater than or equal to the reconstruction error threshold, the parameters of the feature embedding layer, the bidirectional encoding layer and the attention cohesion layer are reversely updated based on the reconstruction error threshold and the comprehensive feature vector set is reprocessed.

7. The fault probability monitoring and control system according to claim 1, characterized in that: The acquisition module includes an acquisition control unit; The acquisition control unit acquires M pending related parameters in the kth cycle to generate a pending related parameter vector, determines the related parameters of the secondary acquisition based on the re-acquisition annotation vector of the kth cycle, and replaces the corresponding pending related parameters with the secondary acquisition results to generate a related parameter vector.

8. The fault probability monitoring and control system according to claim 1, characterized in that: The acquisition module also includes an acquisition verification unit; The acquisition verification unit extracts the mth undetermined related parameter of the kth cycle and traverses the order q m The value set of the mth related parameter fitting order q before the kth period m For each value of the autoregressive model, select the optimal order with the smallest Akaike Information Criterion And the corresponding optimal autoregressive model, substituted into the k-th period before The mth related parameter of the kth period is used to obtain the mth theoretical related parameter of the kth period, and the mth parameter error of the kth period is calculated in combination with the mth undetermined related parameter of the kth period. Based on the comparison result of the mth parameter error of the kth period and the mth parameter error threshold, the mth re-acquisition annotation of the kth period is set, which is organized into the re-acquisition annotation vector of the kth period and fed back to the acquisition control unit.

9. The fault probability monitoring and control system according to claim 8, characterized in that: Determine the optimal order of the mth correlation parameter And the corresponding optimal autoregressive model includes the following steps: Ergodic order q m The value set of , based on the order q m The single value of constructs the variational autoregressive model, that is, the mth related parameter of each period is equal to the q before each period m The weighted sum of the mth correlation parameters of the period; Get q before the kth cycle m The mth correlation parameter of the +1 period is fitted by the least squares method to obtain the q of the variational autoregressive model m autoregressive coefficients; Calculate the order q m Akaike Information Criterion for a single-valued variational autoregressive model; Select the optimal order that minimizes the Akaike Information Criterion And the corresponding optimal autoregressive model is obtained.

10. A method for monitoring and controlling failure probability, characterized in that: The following steps are involved: The undetermined correlation parameter vector is acquired in the kth cycle, and the optimal order of the minimum Akaike Information Criterion is determined based on the mth correlation parameter fitting before the kth cycle. and the corresponding best autoregressive model; Before the kth cycle Substitute the mth related parameter of the kth period into the best autoregressive model to obtain the mth theoretical related parameter of the kth period, combine the mth undetermined related parameter of the kth period to calculate the mth parameter error of the kth period and compare it with the mth parameter error threshold; If it is greater than or equal to the mth parameter error threshold, perform secondary acquisition and use the acquisition result as the mth related parameter of the kth cycle. If it is less than the mth parameter error threshold, use the mth undetermined related parameter of the kth cycle directly as the mth related parameter of the kth cycle to generate a related parameter vector. The window is shifted N times in reverse order starting from the kth period. The time series statistics and cross-correlation coefficient matrix in each window are calculated and flattened into a vector. The vectors are then combined with the relevant parameter vectors of the end period in each window to form a comprehensive feature vector in each window. The core feature vector set is then extracted using the time series autoencoder. Define the state S and use the forward-backward algorithm to maximize the log-likelihood function. Combined with the core feature vector set, the hidden Markov model is trained to obtain the optimal parameter set. Bayesian reasoning is used to calculate the failure probability of the k-th cycle and compare it with the probability threshold to determine whether the k-th cycle is a failure. If it is greater than or equal to the probability threshold, it is determined that a fault has occurred and the equipment is stopped. According to the error between each relevant parameter of the kth cycle and the corresponding standard relevant parameter, it is converted into the control amount of the control device corresponding to each relevant parameter of the kth cycle based on the PID control algorithm and the control is executed. If it is less than the probability threshold, no processing is performed.

Images